Beruflich Dokumente
Kultur Dokumente
Haoyan GUO
School of Computer Science and Technology, Harbin Institute of Technology (Weihai), Weihai, China
Lei Liu
Institute of Physics and Electronic Engineering, TaiShan University, Taian, China
Abstract—Texture can describe a wide variety of surface techniques, which describe texture using structural
characteristics and a key component for human visual primitives and layout [1]. Statistical methods are
perception and plays an important role in image-related commonly used and proved to be effective in texture
applications. This paper proposes a scheme for texture analysis [3, 4, 5]. Methods based on Multi-resolution
image classification using visual perceptual texture features
and Gabor wavelet features. Three new texture features
decompose a texture image into different scales, from
which are proved to be in accordance with human visual which more statistics can be extracted and used to
perceptions are introduced. Usually, Subband statistics describe texture features [7, 8, 11, 12, 13]. These methods
based on Gabor wavelet features are normally used to have been effectively used for solving texture recognition
construct feature vectors for texture image classification. problems.
However, most previous methods make no further analysis Due to the resemblance between multi-resolution
of the decomposed subbands or simply remove most detail filtering techniques (such as Gabor and wavelet
coefficients. The classification algorithms commonly use transform) and human visual process, Gabor and wavelet
many features without consideration of whether the features transform techniques are often used for texture
are effective for discriminating different classes. This may characterization through the analysis of spatial-frequency
produce unnecessary computation burden and even
decrease the retrieval performance. This paper proposes a
content. Many publications showed the effectiveness of
method for selecting effective Gabor wavelet subbands using these techniques for texture analysis, segmentation,
based on feature selection functions. The method can retrieval and classification [7, 8, 11, 12, 13].
discard those subbands that are redundant or may lead to However, most previous methods make no further
wrong classification results. We test our proposed method analysis of the decomposed subbands or simply remove
using the Brodatz texture database, and the experimental most detail coefficients. The classification algorithms
results show the scheme has produced promising results. commonly use many features without consideration of
whether the features are effective for discriminating
Index Terms—Visual Perception Texture Features, Gabor different classes. This may result in unnecessary
wavelet features; Texture image classification; SVM
computation and even decrease the classification
performance.
Psychophysical investigation has shown that the Human
I. INTRODUCTION
Visual System (HVS) does a frequency analysis when we
Texture is an important component of human visual see images [22, 23, 24]. Texture is especially suited for
perception and can be effectively used for characterizing this type of analysis for its intrinsic properties. In this
different image regions. For example, the skin of a zebra paper, we introduce three visual perceptual features,
can be seen as texture with black and white stripe pattern. namely directionality, contrast and coarseness, which
In the past, texture features have been extensively studied have proved to be in accordance with human visual
in the research area of texture image classification and perception in image retrieval experiments [17, 18].
content-based image retrieval, as well as other fields Then, a scheme for texture image classification using
related to pattern analysis [1, 2, 3, 4]. visual perceptual texture features and Gabor wavelet
Basically, texture representation methods can be features is proposed. In contrast to previous work, our
classified into three categories, namely structural, study focuses on reducing the dimensionality of feature
statistical and multi-resolution filtering methods. Typical vectors by discarding those subbands that may be
structure-based methods include morphological and graph irrelevant or redundant, with the aim of achieving better
texture classification performance. We introduce a where: {h(n)} and {g(n)} are low pass filter and high pass
procedure for selecting effective wavelet subbands [16]. {C 0 }
filter, respectively; n is the input signal; N is the
The feature selection functions are used to rank all
subbands, so that the problem of selecting the most length of the input signal; L is the necessary progression.
“useful” subband set reduces to picking up those with Wavelet is commonly used for multi-resolution analysis
highest rankings. Statistics extracted from such subbands and representing image variations at different scales. The
are then used to construct feature vectors. In this way, the computation framework of the wavelet transform of a 2D
dimensionality of feature vectors is reduced. Experiments signal involves recursive filtering and sub-sampling, as
show the selected subbands are effective for shown in Fig.1. To clarify the use of symbols, we briefly
discriminating different textures, as the classification describe the general process of wavelet transforms and
performance is much improved. the notations of subbands of wavelet decomposition. At
The rest of the paper is organized as follows. In section each level, there are three detail images. Following [7],
2, we introduce the visual perceptual texture features. we denote the detail images (subband) as LH (contains
Section 3 briefly describes Gabor wavelet decomposition. the high frequency horizontal information), HL (contains
In section 4, a feature selection algorithm is introduced. the high frequency vertical information), and HH
In section 5, experimental results are presented in section (contains the high frequency diagonal information). The
5. Finally, we conclude the paper in section 6. decomposition/transform also produces one
II. NEW TEXTURE FEATURES DERIVING FROM approximation image, denoted by LL, containing the low
WAVELET TRANSFORM frequency information. The wavelet transform can
recursively decompose the LL band. Since 2 level
Wavelet transform has been widely used in image wavelet decomposition yields 6 detail images, we use
processing and shows tremendous advantages over LH1, HL1, HH1, LH2, HL2, HH2, and an additional
Fourier transform. Wavelet transform is a multi- approximation image LL2 to denote all the subband
resolution analysis that represents image variations at images.
different scales [6, 7, 8]. According to the definition of
wavelet, a wavelet is an oscillating and attenuated
function and its integrals equal to zero. It is a
mathematical function useful in digital signal and image LL2 HL2
processing. HL1
Given f(x) is a one-dimensional input signal, a 1-
Ddiscrete wavelet transform is defined as:
LH2 HH2
φ jk ( x) = 2 − j / 2 φ (2 − j x − k ) ,
( j = 1,2,..., L, k = 0,1,..., N 2 j − 1) .
200
100
-200
Fig.2 Three textures from Brodatz database.
-300
M N
1
DirH =
MN
∑∑ p
j =1 k =1
jk
2
, The vertical directionality is defined as :
M N
1
Where M, N are the sizes of HL subband.
DirV =
MN
∑∑q
j =1 k =1
jk
2
,
Fig. 3 (a) is the D1 texture from the Brodatz database and
(b) shows the 10th row’s coefficient convolution result of Where M, N are the sizes of HL subband.
HL1 subband .
B. Contrast z −1 z −1
E z ,v ( s , t ) =| A z ( s , t + 2 ) − Az ( s, t − 2 )|
For the purpose of denoting the change of grey levels .
in a texture, contrast is commonly defined for each pixel
as an estimate of the local variation in a neighborhood. At each point, the value of Z that maximizes E in
The calculation of the contrast is implemented in the either direction is used to set the best size:
wavelet decompose approximation subband, denoted as
LL, which contains the low frequency and reflecting the z
S best ( s , t ) = 2
global information of the texture. ;
Given x ( j , k ) as the coefficient of LL subband, where j
and k represent the row and column values of the The global coarseness is calculated by averaging S best
subband images, respectively, let the pixel at the point (j, over the entire LL subband:
V. EXPERIMENTAL RESULTS
{
K ( x, xi ) = exp − x − xi / γ 2
2
};
This is a radial basis classifier. For most pattern
recognition applications, using different parameters may
have more effects on the correct recognition rate than
using different kernel functions [14, 15].The RBF kernel
can be used in applications related to human vision
characteristics such as license plate recognition.
The SVM algorithm, using the RBF kernel functions,
employed for classification was as described in the above
section. The experiments results for various experimental
training samples are listed in Table 1. Every sample was
divided into 24 sub-images to perform the classification.
For every class, we random selecting 14 textures used for
training and the rest 10 textures for testing. Another
experiment is 18 textures used for training per class, and
the rest 6 textures were used for testing.
As the Table 1 shows, the method using feature selection
algorithm outperforms the method using all the Gabor
wavelet subbands features and visual perceptual texture
features. The feature selection algorithm successfully
selected effective Gabor wavelet subbands that can
improve correct classification rates. Our experiments
show that if we do not carried out the subband selection
procedure, although Gabor wavelet can capture more
orientation information of the sample texture than
selected effective subbands, the classification
performance is worse. We deduce that many subbands are
not effective and discriminating or redundant when used
as texture features. After the subband selection procedure,
although the dimensionality of the texture feature vector
is lower, the classification accuracy is raised. The average
accuracy produced in different ways is shown in Table 1.
ACKNOWLEDGMENT
This work is supported by the National Natural Science
Foundation of China and Chinese Academic of Sciences
joint fund on astronomy under project No. 10778619.
D97 REFERENCES
[1] Long F., Zhang H.J., and Feng D.D. (2003)
"Fundamentals of content-based image retrieval" in
Multimedia Information Retrieval and Management-
Technological Fundamentals and Applications, D. Feng,
(a) (b) (c) (d) W.C. Siu, and H.J. Zhang (Eds.), Springer, 2003.
[2] Howarth, P., Ruger, S.: Evaluation of texture features for
content-based image. retrieval. In: Proceedings of the
International Conference on Image and Video Retrieval,
Springer-Verlag (2004) 326–324
[3] H. Tamura, S. Mori, and T. Yamawaki, "Texture features
corresponding to visual perception," IEEE Trans. On
Systems, Man, and Cybernetics, vol. Smc-8, No. 6, June
1978.
[4] F. Liu, and R. W. Picard, "Periodicity, directionality, and
randomness: Wold features for image modeling and
retrieval," IEEE Trans. on Pattern Analysis and Machine [18] Muwei, Jian, Junyu Dong, “New Perceptual Texture
Learning, Vol. 18, No. 7,July 1996. Features Based on Wavelet Transform”. International
[5] R. W. Picard, T. Kabir, and F. Liu, "Real-time recognition Journal of Computer & Information Science (IJCIS).
with the entire Brodatz texture database,"Proc. IEEE Int. Volume 9, Number 1, pp 11
Conf. on Computer Vision and Pattern Recognition, pp. [19] Georgeson, M. A., “Spatial Fourier Analysis and Human
638-639, New York, June .1993. Vision,” Chapter 2, in Tutorial Essays in Psychology, A
[6] I. Daubechies, "The wavelet transform, time-frequency Guide to Recent Advances, N. S. Sutherland (ed.), vol 2,
localization and signal analysis," IEEE Trans. on Lawrence Earlbaum Associates, Hillsdale, N.J., 1979.
Information Theory,Vol.36,pp.961-1005,1990 [20] Battiato, S. Gallo, G. Nicotra, S. Perceptive visual texture
[7] B. S. Manjunath, and W. Y. Ma, "Texture features for classification and retrieval. 12th International Conference
browsing and retrieval of image data," IEEE Trans. on on Image Analysis and Processing pp. 524- 529. Sept.
Pattern Analysis and Machine Intelligence, Vol. 18, No. 8, 2003
pp. 837-842, Aug. 1996. [21] P. Brodatz, Textures: A Photographic Album for Artists&
[8] Manjunath, B. Wu, P. Newsam, S. and Shin, H. “A Texture Designers, Dover, New York, 1966.
Descriptor for Browsing and Similarity Retrieval.” Journal [22] Georgeson, M. A., “Spatial Fourier Analysis and Human
of Signal Processing: Image Communication, 16(1), pp. Vision,” Chapter 2, in Tutorial Essays in Psychology, A
3343, September 2000. Guide to Recent Advances, N. S. Sutherland (ed.), vol 2,
[9] Abhir H. Bhalerao and Nasir M. Rajpoot.” Discriminant Lawrence Earlbaum Associates, Hillsdale, N.J., 1979.
Feature Selection for Texture Classification”.In proc. [23] Devalois, R.L., D.G. Albrecht, and L.G. Thorell, “Spatial -
British Machine Vision Conference ,sep.2003 frequency selectivity of cells in acaque visual cortex,”
[10] O. Boz. Feature Subset Selection by Using Sorted Feature Vision Research, 22, pp. 545-559, 1982.
Relevance. In Proc. Intl. Conf. on Machine Learning and [24] Caelli, T., Visual Perception, Pergamon Press, 1981.
Applications, June 2002.
[11] Simona E. Grigorescu, Nicolai Petkov, and Peter Kruizinga
“Comparison of Texture Features Based on Gabor Filters”
IEEE Transactions on image processing, Vol. 11, no. 10,
OCTOBER 2002
[12] J. Smith and S.-F. Chang, “Transform features for texture
classification and discrimination in large image database”. Muwei Jian is a lecturer in the
IEEE Intl. Conf. on Image Proc., 1994. School of Space Science and Physics,
[13] Y. Ma and B. Manjunath, “Texture features of wavelet Shandong University at Weihai, China.
transform features for texture image annotation”, IEEE Mr. Jian holds 3 granted patents and
Intl. Conf. on Image Proc., 1995. has published more than 20 journal and
[14] Vapnik V N. Statistical Learning Theory [M]. New York: international conference papers
John Wiley & Sons Inc, 1998. (indexed by EI and ISTP). His research
[15] XueGong Zhang. Introduction to statistical learning theory interests include image processing,
and support vector machines[J]. Acta Automatica wavelet analysis, multimedia analysis
[16] Dong, Junyu; Jian, Muwei; Gao, Dawei; Wang, Shengke. and computer vision. He has investigated issues in content-
"Reducing the dimensionality of feature vectors for texture based image/multimedia retrieval, 3D texture synthesis, editing
image retrieval based on wavelet decomposition". and synthesis of 3D surface texture, texture classification and
Proceedings - SNPD 2007: Eighth ACIS International image fusion etc.
Conference on Software Engineering, Artificial
Intelligence, Networking, and Parallel/Distributed
Computing, v 1, 2007, p758-763..
[17] Muwei, Jian, Dong, Junyu; Gao, Dawei; Liang, Zuojuan Haoyan GUO is a lecturer in the School of Computer Science
"New texture features based on wavelet transform and Technology, Harbin Institute of Technology (Weihai),
coinciding with human visual perception". Proceedings- Weihai, China.
SNPD 2007: Eighth ACIS International Conference on
Software Engineering, Artificial Intelligence, Networking,
and Parallel/Distributed Computing, v 1, 2007, p 369-
373. Lei Liu is a lecturer in the Institute of Physics and
Electronic Engineering, TaiShan University, Taian, China.